Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual transfer and Beyond

[flrngel]

http://arxiv.org/abs/1812.10464

Notes

• Model learn joint multilingual sentence representations

Pre-training techniques

using Moses tools

• punctuation normalization
• non printing chars removal
• Tokenization

Paper summary

• this paper introduces new dataset using tatoeba for language similarity search
  • covers 122 languages
  • consists 1,000 English-aligned sentence pairs for each language
• enables to use sentence embedding with English annotated only, and transfer to other languages without any modifications